An efficient deep learning network with orthogonal softmax layer for automatic detection of tuberculosis

Abstract

Tuberculosis (TB) is a persistent bacterial lung infection that affects the lungs severely. It is one of the fatal diseases among the top ten leading causes of death. Thus, early and accurate diagnosis of TB precisely has vital role in saving lives. In this work, a novel Deep Learning Network with Orthogonal SoftMax Layer (OSL) is proposed for the detection of Tuberculosis (TB) that yields admirable performance in small medical datasets. The integrated merits of deep learning-based efficient feature extraction and OSL-based classification boost the classification performance. OSL enhances discriminative feature learning capability by maintaining orthogonality among weight vectors in this layer. It is also responsible for reducing the co-adaption among parameters and making optimization easier. In addition, data augmentation is done in the training stage using a customized version of RandAugment to train the network properly, even with small medical datasets. The experimental results depict the proposed deep learning network outperforms others with the best 98.67% accuracy, 98.99% precision, 98.33% sensitivity, 99.00% specificity, and 0.9866 F1 score.